Scientific Machine Learning for Drug-Target Interaction Predictions

Accurately predicting drug-target interactions (DTIs) is crucial for expediting the drug development process, saving time, resources, and enhancing drug safety. While traditional methods either lack accuracy or are computationally expensive, existing machine learning models in this domain often suffer from a lack of generalization due to the complex nature of molecular interactions. Here, we show that integrating physics-based knowledge into a data-driven model significantly improves prediction accuracy and model interpretability.

Our approach involves equations derived from chemical interaction physics and using a gated graph attention network to capture the
intricacies of atomic interactions. We demonstrate that our model not only predicts the total binding affinity of drug compounds to target proteins but also calculates pairwise atom-atom energies, offering a detailed understanding of molecular interactions. This granular analysis is vital for optimizing drug efficacy and guiding drug discovery efforts.

Our results indicate better performance compared to traditional methods, with enhanced generalization and robustness against input variations, thus providing a promising direction for more targeted and effective drug design strategies.